assessing avian habitat fragmentation in urban areas of

Author's personal copy

Landscape and Urban Planning 95 (2010) 54–60

Contents lists available at ScienceDirect

Landscape and Urban Planning

journa l homepage: www.e lsev ier .com/ locate / landurbplan

Assessing avian habitat fragmentation in urban areas of Hong Kong (Kowloon)at high spatial resolution using spectral unmixing

Janet Elizabeth Nichola,∗, Man Sing Wonga, Richard Corlettb, Douglas W. Nichola

a Department of Land Surveying and Geo-Informatics, The Hong Kong Polytechnic University, Hong Kongb Department of Biological Sciences, National University of Singapore, 14 Science Drive 4, Singapore 117543, Singapore

a r t i c l e i n f o

Article history:Received 21 July 2008Received in revised form15 November 2009Accepted 4 December 2009Available online 8 January 2010

Keywords:Ecological corridorsLeast cost path analysisLinear spectral unmixingSatellite imagesUrban vegetation

a b s t r a c t

The fragmentation, isolation and sparseness of vegetation in urban areas gives small patches of vegetationenhanced ecological value as habitat islands, compared with rural areas. Habitats as small as an individualtree may provide important landscape linkages across a densely urbanized city. Thus there is a need formicro-scale inventories of whole cities, which should incorporate not just biomass, but also vegetationof different life form, to support different habitat structural requirements. A combination of fine resolu-tion multispectral satellite images from Ikonos, with the image processing technique of Linear SpectralUnmixing (LSU) permits the identification of different types of urban vegetation at sub-pixel level. Anyfractional amount of grass and/or trees respectively within each 4 m Ikonos pixel can be identified.

To evaluate such fine scale inventory, least cost path (LCP) analysis was performed using pixel fractionsrepresenting micro-scale tree habitats. This study adopts an innovative approach by allocating variableweightings to the vegetation fraction amounts within each pixel, rather than to whole pixels. The resultis a fuzzy friction surface, which constitutes a very high-resolution database for input to fuzzy queryingand decision-making. The friction values represent species’ preferences or tolerance levels, and may bevaried according to the fraction amounts within a pixel. Automated mapping of least cost pathways overdifferent friction surfaces produced different routes across the study area, the densely urbanized KowloonPeninsula in Hong Kong. Comparison of the results with field data of bird sightings indicates the need forhigh detail in urban ecological analysis.

© 2009 Elsevier B.V. All rights reserved.

1. Introduction

The fragmented nature of wildlife habitats in modern cities(Bolger et al., 2000; Park and Lee, 2000; Huste and Boulinier, 2007)has prohibited sufficiently detailed inventories over whole cities,for comprehensive landscape ecological analysis. Not enough isknown about how wildlife interacts with whole urban landscapes(Blair, 1996; Fernandez-Zuricic and Jokimaki, 2001; Melles et al.,2003), and Savard et al. (2000) recommend urban planners toinventory resources within a city at scales ranging from individualplants to the entire city. Until recently, field survey (e.g. Blair, 1996;Jokimaki, 1999) and manual air photo interpretation (e.g. Porter etal., 2001; Huste and Boulinier, 2007) have been the norm, and mostresearch has concentrated on designated green areas such as parks(Jokimaki, 1999; Lock, 2000), pre-selected plots (Whitcomb, 1977;Opdam et al., 1984), or localized transects across urban–rural gra-

∗ Corresponding author. Tel.: +852 2766 5952; fax: +852 2330 2994.E-mail addresses: [email protected] (J.E. Nichol),

[email protected] (M.S. Wong), [email protected] (R. Corlett),[email protected] (D.W. Nichol).

dients (Blair, 1996; Clergeau et al., 1998; Lock, 2000), with only afew city-wide studies (Melles et al., 2003). Field surveys which havecovered whole cities, in Hong Kong (Jim, 2004) and New York (Smalland Lu, 2006), included only trees and were extremely manpowerintensive. However Sodhi et al. (1999), in investigating bird usageof park connectors in urban Singapore observed that the designatedroutes with the most diverse bird communities were bordered byinformal green space. This supports the (now long-standing) recog-nition that many urban unmanaged areas and informal green spaceare rich in species (Sukopp and Weiler, 1988; Wee and Corlett,1986; Huste and Boulinier, 2007; Bino et al., 2008) and suggeststhat conserving recreational spaces and green areas, while ignor-ing non-designated vegetation is not the best approach. Thus analternative for urban ecological surveys is to consider all exist-ing habitats both natural and man-made, if extensive and detailedvegetation maps at life form level are available. However sinceurban vegetation is usually fragmented, previous low and mediumresolution satellite sensors have been unable to resolve micro-habitats in densely urbanized areas. Such micro-habitats may haveenhanced importance due to the oasis effect (Macarthur and Wil-son, 1963; Wiens, 1994; Fernandez-Zuricic and Jokimaki, 2001)and their role in connectivity (Belisle, 2005). ‘Fragmentation’ is

0169-2046/$ – see front matter © 2009 Elsevier B.V. All rights reserved.doi:10.1016/j.landurbplan.2009.12.002


J.E. Nichol et al. / Landscape and Urban Planning 95 (2010) 54–60 55

used in this study to refer to the heterogeneous and discontinu-ous patchwork of urban vegetation (which may include single, orrows of street trees, shrubs and grassy road verges, of both nativeand introduced species), rather than Fahrig’s (2003) definition ofpatches derived from the breaking apart of a once continuous coverof natural vegetation accompanied by loss of habitat area.

This paper demonstrates that maps of vegetation life form pro-duced from sub-pixel analysis of fine resolution satellite imagesare more effective and relevant for urban ecological analysis, thanthose produced from the previous generation of moderate reso-lution sensors, or from traditional ‘per pixel’ classifiers applied tofine resolution images. The ‘soft’ classification technique of LinearSpectral Unmixing (LSU) identifies the fractions (percentages) ofgrass and/or trees within a pixel (Nichol and Wong, 2007) and isthus able to generate very high-resolution fuzzy vegetation mapscovering a whole city. These can be input directly as fuzzy sur-faces (Lodwick, 2008) to spatial analysis routines for ecologicallandscape assessment, and automated least cost path (LCP) anal-ysis (Berry, 1993; Ares et al., 2007) to depict functional pathways isgiven as an example in this study. The technique described pro-vides a high-resolution raster database for landscape ecologicalmodeling and hypothesis testing, which appears to be more rele-vant for understanding wildlife interactions, than previous scales ofenquiry.

2. Techniques for urban vegetation survey

Sub-pixel analysis of fine resolution images such as Ikonos(Nichol and Wong, 2006) or Quickbird (Small and Lu, 2006) per-mits highly detailed and accurate vegetation mapping over a wholecity. The image processing technique of Linear Spectral Unmixing(LSU) (Adams et al., 1986; Ridd, 1995; Phinn et al., 2002; Nicholand Wong, 2007) is a soft classifier, which avoids the mixed pixelproblem of hard classifiers such as maximum likelihood (MLC) bycomputing the proportions of cover types (endmembers) within apixel, and does not allocate pixels to a single class. The outputsfrom LSU are fraction images whose pixel values represent thepercentage of a given cover type within a pixel. Nichol and Wong(2007) demonstrated that trees and grassy surfaces can be recog-nized and mapped as distinct pixel fractions using the uniquely highspatial, spectral and radiometric properties of the Ikonos satellitesensor. Where the pixel size is 4 m × 4 m, as with Ikonos, a 20%tree fraction represents one or more micro-habitats totaling 3.2 m2

in size within that 4 m × 4 m area. Furthermore, since the fractionimages are fuzzy datasets they are amenable to fuzzy querying anddecision-making, as demonstrated in this paper.

3. The study area

The study area comprises the Kowloon Peninsula of Hong Kong(Fig. 1), which is approximately 160 km2 in extent, and one of themost densely urbanised areas in the world. The most densely builtareas are devoid of vegetation, and elsewhere street planting isseverely restricted by lack of space (Jim, 2004). The urban areais surrounded by mountainous regions supporting evergreen sec-ondary forests and the urban boundary is usually abrupt, at thefoot of steep slopes. Within Kowloon, the main vegetation com-prises street trees, local (mainly treed) parks, institutional managedgrassland and a few steep inselbergs supporting remnant secondaryforest. The avifauna of Hong Kong is considered to be species rich(Lock, 2000), with a total of 450 species recorded in its 1098 km2

land area. Lock (2000) recorded 20% of these in the 14 ha. KowloonPark which is approximately 6 km distant from the nearest coun-tryside across land. Although she found the avifauna of urban parksin Hong Kong to be more similar to cultivated land than to forest,

the Great tit (Parus major) which is predominantly a forest bird inHong Kong, was observed in five out of six urban parks.

4. Methods

4.1. Vegetation mapping

Separate images representing grass and tree pixel fractions werederived by LSU from a December 2001 Ikonos image of 4 m spatialresolution with 4 bands (RGB/IR) (Nichol and Wong, 2007) coveringthe whole of urban Kowloon. The procedures involve the deriva-tion of pure spectral signatures from each of the main land covertypes, known as endmembers. These were obtained by selectingpure, unmixed pixels from the image. Lastly the fractional covertypes within each pixel were obtained by the spectral unmixingprocedure which is based on the solution of a set of least squareequations for n endmembers and n bands. For our study, four end-members, namely grass, trees, high albedo and low albedo surfaceswere selected, and the four Ikonos bands plus an additional NDVIband were used. The end product used in the present study is animage whose pixel values represent the percentage (fraction) oftrees within the pixel. Tree cover maps for the study area were alsoobtained using a standard MLC (whole pixel) classification of the4 m Ikonos and 20 m SPOT images. This uses training area statisticsto allocate each whole pixel to one land cover type. There was noattempt to differentiate between trees and shrubs as distinct lifeforms in this study since field measurement with a multispectralradiometer showed no significant difference between the spectralreflectance of trees and shrubs. However this may not be a greatdisadvantage for habitat evaluation since Lock (2000) found thatthere was significant overlap in the species composition of birdsfound in urban tree and shrub habitats in Hong Kong, comparedwith distinctly different species in grassland.

Since the LSU technique assumes that the contribution of the dif-ferent cover types to the total pixel reflectance is linearly relatedto their area, some error may be generated, whereby the fractionalproportions of the endmembers may not equal 1. Accuracy assess-ment for LSU is difficult because pixel unmixing does not producehard boundaries of land cover types whose position and size can bemeasured. Furthermore, overlay onto a reference dataset of knownaccuracy is limited by the accuracy of the geometric correction pro-cess which is usually only ±0.5 pixel. In this study this problemwas mitigated by using the control points obtained from orthorec-tification of the 1 m Ikonos panchromatic band, for rectifying the4 m resolution multispectral image. Thus accuracy assessment, byoverlay of 70 Ikonos LSU test pixels onto a high (0.8 m) resolutioncolour orthophoto minimized errors of co-registration. A maxi-mum error of ±0.3 m over the 16 m2 area of an Ikonos pixel wasobtained, which corresponds to ±2.5 m2, or 15%. The position ofthe Ikonos test pixels on the orthophoto was further verified visu-ally with reference to well-defined features such as plot corners,buildings and intersections. Within each 4 m × 4 m Ikonos pixel onthe orthophoto the grass and trees were visually interpreted, andtheir areas screen digitized and measured.

4.2. Least cost path (LCP) analysis

In order to evaluate whether such a high resolution database isnecessary for common urban ecological applications, LCP Analysiswas undertaken to compare pathways generated from both fine andmoderate resolution images, and from both hard and soft classifica-tion techniques. The pathways represent routes taken by birds withdiffering habitat requirements to reach Kowloon Park. The LCP bal-ances habitat suitability, degree of connectivity between startingand end points, and minimum Euclidean distance by considering


56 J.E. Nichol et al. / Landscape and Urban Planning 95 (2010) 54–60

Fig. 1. Grass and tree fraction image derived from LSU, showing grass as red and tree as turquoise. The percentage of grass and/or trees in each pixel is represented by therespective colour brightness, and pixels containing both grass and trees appear magenta.

the cumulative pixel values at each grid point. These considerationsare based on previous research which indicates that both local andlandscape level factors influence bird abundance and species rich-ness (Blair, 1996; Melles et al., 2003), and that birds will travel muchfarther across wooded habitats than across small gaps (Desrochersand Hannon, 1997). Thus LCP analysis identifies the path of leastresistance across a cost surface, where costs are usually based onthe allocation of weightings to whole pixels representing land covertypes. This study adopts an innovative approach with the LSU softclassifier, by treating the LSU fractions as a fuzzy dataset wherebyweightings are allocated according to the percentage of a singleland cover type (in this case tree cover) within a 16 m × 16 m pixel.

Tree fraction images were generated by LSU from 4 m Ikonosmultispectral images and imported to IDRISI v.14.02 (Clark Labs,Worcester, MA, USA). The pixel fraction amounts were divided

into 5 equal classes for query purposes, and each class was givena weight representing the perceived degree of difficulty for birdsto travel across the varying fractions of tree cover (Table 1). Thisfriction surface was then input to the IDRISI COST module whichcomputes a cost surface by considering both the friction value anddistance from a defined target. Three different weighting schemeswere devised to test for sensitivity to differing fractions of tree coverwithin a pixel (Table 1). Based on an ending point at the largesturban park, Kowloon Park, and starting at the nearest continu-ous area of dense forest, the PATHWAY module (Eastman, 2006)was used to compute three LCPs for the three weighting schemes.For comparison purposes, two additional friction surfaces and LCPswere created from less detailed tree cover maps derived from themaximum likelihood (whole pixel) classifier of (i) the 4 m Ikonosimage and (ii) a 20 m resolution SPOT image.



Table 1Left hand part of the table gives the friction values allocated to different tree fraction classes, for input to least cost path mapping. The bird data in the right of the table arederived from four field visits to each route, and numbers in brackets represent computation without habitat generalists (feral pigeon (Columbus livia), tree sparrow (Passermontanus) and black kite (Milvus lineatus)).

Fraction of treecover

Allocatedfriction value

Description ofroute suitability

Density:birds/ha.

Number ofspecies

Insect ivorea

densityShannondiversity index

Equita bilityindex

Route A. Route length: 7454 m (N-S), 3562 m (E-W)0.8–1 5

Recognises only dense trees. Lowertree fractions equal to no trees

North-South path0.6–0.8 50 9.4 (6.6) 15 (12) 1.4 2.26 0.790.4–0.6 500.2–0.4 50 East-West path0–0.2 50 8.1 (3.7) 8 (5) 0 1.29 0.62

Route B. Route length: 7358 m (N-S), 3416 m (E-W)0.8–1 5

Prefers denser trees but can traversenon-tree as last resort

North-South path0.6–0.8 10 9.1 (6.7) 14 (11) 0.85 2.27 0.800.4–0.6 200.2–0.4 50 East-West path0–0.2 100 12.5 (8) 11 (8) 3.6 1.81 0.75

Route C. Route length: 6473 m (N-S), 3209 (E-W)0.8–1 10

All tree fractions equally suitable;avoids gaps

North-South path0.6–0.8 10 16.8 (9) 19 (16) 1.5 2.37 0.800.4–0.6 100.2–0.4 10 East-West path0–0.2 100 21 (11) 12 (9) 1.4 1.89 0.76

a Includes common tailorbird (Orthotomos storius), great tit (Parus major), and yellow-browed warbler (Phylloscopus inornatus).

For the purpose of the study, birds are assumed to choose theleast cost (optimum) path, since they would encounter fewer haz-ards, would spend less time in traveling, and would travel throughhabitat with higher probability of containing food and cover. It ishypothesized that if the allocation of weightings to tree fractionsof less than 100% of a pixel were to result in significantly differ-ent and shorter routes than for the MLC-based routes, then treepatches smaller than the 16 m2 of an Ikonos pixel must be impor-tant for providing habitat connectivity across the urban landscape.Furthermore, if the biological relevance of the different routes wasshown to differ, with birds showing a preference for shorter routeshaving more fragmented tree cover, this would support the needfor such high resolution mapping in urban areas. This hypothesisis supported by previous work indicating that species richness ismore related to connectivity and than to habitat area (Jokimaki,1999; Molainen and Hanski, 2006).

The routes generated by the three weighting schemes for theLSU tree fraction image traverse Kowloon in a generally north-south direction, and are as follows:

(i) In Route A, high friction values were given to all pixels exceptthose with the highest tree fraction value of 0.8–1 (column 2,Table 1), to simulate a route favoured by birds which require athick tree cover and cannot tolerate sparse, or no trees. This isthe longest LSU route, at 7454 m as it is unable to utilise frag-mented tree cover, and follows dense tree patches approachingthe size of one 4 m × 4 m pixel or larger. It traverses betweenlarge patches across the shortest urban distance, not recognis-ing any smaller intervening tree fractions.

(ii) For Route B, the friction weights allocated to the tree fractionswere set to increase gradually as the percentage (fraction) oftree cover decreased. This route considers birds which preferdense tree cover but which could also tolerate lesser amounts,albeit as a lower preference. Route B follows the shortest routeC in the southern half but, having more stringent requirementsfor tree cover, takes a more circuitous path in the northern half,where it is less able to exploit the fragmented tree cover of theresidential areas. It thus diverges to follow parks and largertree patches.

(iii) Route C, where the friction values of all pixels containing anytree fraction are set equal, but lower than those with no tree,

represents a requirement for any amount of tree cover alongthe route but the patch size is not a limiting factor. This is theshortest and most direct route, at 5918 m, as the requirement isfor any tree cover regardless of the fraction amount. This routecomprises both low density residential development with frag-mented tree cover, as well as vegetated linear corridors alongthe Kowloon-Canton Rail (KCR) line, and a double row of oldtrees along Wylie Road.

In order to repeat the procedures and verify the robust-ness of the findings, three additional routes with the sameweightings (A, B and C), but in an east-west direction, weregenerated.

4.3. Field data collection

In order to investigate whether the different LCPs represent realdifferences in their functionality as bird habitats, fieldwork wasundertaken in January and February 2006. Since Kowloon is fullydeveloped, very little vegetation change is known to have takenplace in the 4+ years since image acquisition. A slow walk wasundertaken along the differing sections of the routes mapped byLSU, recording with binoculars the number and species of birdsseen within a 20 m strip on either side of the route. This distancewas chosen because DeGraaf et al. (1991) noted that more than 70%of urban birds can be seen within a distance of 20 m, but this falls to40% at 30 m and only 20% at 40 m. The walks were repeated on 4 sep-arate occasions, and the results averaged. Bird density and speciesnumbers were computed both with and without three abundantspecies which appear not to require green areas: namely feralpigeon (Columbus livia), tree sparrow (Passer montanus) and blackkite (Milvus lineatus), which will be referred to as habitat general-ists. The density of insectivores was also noted, since insectivoresindicate the development of a more complex food chain, and aredeemed to be environmentally sensitive species. They comprise alarger proportion of the birds in rural areas (Beissinger and Osborne,1982; DeGraaf and Wentworth, 1991) and insufficient, or planted,as opposed to native vegetation may be unable to support them.The insectivores observed in the study include common tailorbird(Orthotomos sutorius), great tit (P. major) and yellow-browed war-bler (Phylloscopus inornatus).



Fig. 2. Histograms showing frequency of pixels containing each of five fractionclasses of grass and trees. For example most pixels containing tree cover are in thelowest class, which has only 0–20% of tree cover (i.e. less than 3.2 m2), indicatingextreme fragmentation of tree habitats.

5. Results

5.1. Spectral unmixing

Grass and tree fraction images resulting from LSU are shownin Fig. 1 as a colour composite, with grass displayed as red andtrees as blue. Darker hues represent lesser fractions and magentahues would represent pixels having both grass and trees in vary-ing proportions, although this is uncommon in Hong Kong. Highaccuracy was demonstrated by the observed adjusted correlationcoefficients (R2) of 0.94 for grass and 0.96 for tree fractions, whencompared with a high (0.3 m) resolution digital orthophoto (Nicholand Wong, 2007). The pixel values of the images resulting from LSUrepresent percentages of grass and/or tree accordingly, within eachpixel. The extreme fragmentation of both grass and tree ecosystemsin the study area is demonstrated by querying of the pixel values(Fig. 2), which showed that for 84% of pixels containing grass, and72% of pixels containing trees, the grass and tree fractions occupiedless than 60% of the pixel (smaller than 10 m2), respectively. Thisindicates the difficulty of mapping urban vegetation over a wholecity by field survey or manual air photo interpretation, as well as thelikelihood of large error from hard (whole pixel) classifiers such asMLC. The adjusted correlation coefficient for MLC in the same studywas 0.77.

5.2. Least cost path analysis from LSU

The three LCPs (routes A, B and C in Fig. 3) were significantlydifferent over most of the routes, for both the north-south (N-S)and the east-west (E-W) routes. Because the N-S and E-W routeswere very similar in terms of relationships between tree densityand bird populations, discussion will be limited to the N-S routesunless stated, but results for both are given in Table 1.

Field data collected along the three routes indicated, some-what surprisingly, that route C, the shortest, most direct route notrequiring large tree patches, had a 27% higher density of birds, and32% more species than the other two routes. The higher densityand species diversity apply both with and without the three habi-tat generalists not requiring green areas, and which were presentin large numbers. The Shannon Diversity Index, which accountsfor both abundance and evenness of species present, was alsosomewhat higher for route C. Shannon’s Equitability Index, whosevalues lie between 0 and 1, with 1 representing evenness typi-cal of natural habitats, was not significantly different among the

Fig. 3. False colour Ikonos image of Kowloon showing the five north-south routes,and 3 east-west routes derived from the IDRISI PATH module. Route C in both casesis the shortest LSU route.

three routes (Table 1). Insectivore density was also not significantlydifferent.

Visual field observations indicated two possible reasons for thegreater richness of birds along route C (the shortest route, withsmall but more abundant tree patches). Firstly the parks includedin the longer routes A and B were often on steep and shady slopeswhere planting of non-native species has taken place. These includeAcacia confusa, Albizia lebbeck, and Casuarina equisetifolia whichprovide no fruit for birds and produce ecologically impoverishedhabitats due to tough and un-nutritious leaf litter. Secondly smalltree patches as along route C, or even single isolated trees of oldnative species may be very rich in bird numbers as well as species.For example on several occasion old, isolated rubber trees (Ficuselastica) contained more than 30 birds including Red-whiskeredbulbuls (Pycnonotus jocosus), Chinese bulbuls (Pycnonotus synensis)and Asian tree sparrows (P. montanus). Along all routes, most birdswere observed in trees and appeared to be either foraging, nestingor singing, with little evidence of the routes being used as flyways.Fieldwork for the E-W route was done in summer, as opposed towinter for the N-S route, and while bird density was higher, fewerspecies were recorded due to the absence of winter visitors suchas the Hoopoe (Upupa epops), Yellow-browed warbler (P. inorna-tus), and White wagtail (Motacilla alba). However, the ratio between



routes A, B and C, in terms of bird density, and species number anddiversity was remarkably similar, ie. for both N-S and E-W routes, Aand B had approximately half of the bird numbers, and substantiallyfewer species than route C (Table 1).

5.3. Least cost path analysis from whole pixel classifier (MLC)

The two routes derived from the MLC (whole pixel) classifierfor Ikonos at 4 m resolution and SPOT at 20 m resolution respec-tively both appear less able to traverse the urban area than thethree Ikonos LSU routes, since they diverge westwards along theforest margin before turning south in the direction of the end point.Although the Ikonos MLC route is identical to the LSU routes inthe lower half, the SPOT route differs along its whole length fromall others. Since only whole SPOT pixels classified as trees by MLCcan be given lower friction values, the route must follow large treepatches of at least 20 m × 20 m. Some of these route differences arealso probably due to the mixed pixel problem of whole pixel clas-sifiers, since for SPOT, accuracy for the tree class was only 53%. Theonly section common to all five N-S paths is a 700 m section alongthe KCR railway line, which passes through a commercial districtdevoid of trees. This may be identified as a key connecting corridorlinking habitats in the northern and southern sections of the studyarea.

6. Discussion

The generation of significantly different LCPs from different LSUfraction weightings indicates that tree patches smaller than the4 m resolution of Ikonos play an important role in the connec-tivity of tree habitats in the study area. The observation that theshortest, most fragmented route contained both the highest birdnumbers and greatest species diversity (excluding habitat gen-eralists) may suggest that such small tree patches may play animportant role in supporting birds either individually as habitatislands or due to their enhanced importance as habitat connectors.This finding that small tree patches may be at least as importantas larger patches conflicts with the majority of previous studieswhich show a decline in species diversity with decreasing patchsize. However, the patches measured in these studies have usu-ally resulted from the fragmentation of a once continuous naturalhabitat (Fahrig, 2003) rather than from the planting and landscap-ing common in urban areas. Moreover, most of these (e.g. Parkand Lee, 2000; Donnelly and Marzlufff, 2004; Marzluff, 2005) haveinvestigated landscape parameters at grain sizes an order of mag-nitude higher than in our study, over larger study areas and overmuch less densely urbanized landscapes than that of Hong Kong.For example Donnelly and Marzlufff (2004) using Landsat with30 m resolution identified the threshold size for conserving for-est bird species diversity as 42 ha, which is three times larger thanthe largest urban park in the present study. Bino et al. (2008) alsousing Landsat suggested that the size of green area needed to main-tain bird species diversity was 15 ha. On the other hand, resultsfrom more detailed mapping using manual air photointerpretationin a highly urbanized and fragmented landscape in the Paris sub-urbs (Huste and Boulinier, 2007) indicate that for urban-adaptedbirds the patch size near the city centre did not control bird diver-sity, since extinctions were balanced by immigration from nearbypatches, and Fahrig (2003) describes many instances where smallpatches in a landscape have been beneficial to biodiversity. Arthro-pods which are less capable than birds of traversing urban surfacesmay also favour proximity and contiguity over habitat size, andwell-connected tree patches in highly urbanized landscapes maypromote this, along with well-developed predator–prey relation-ships, although the incidence of insectivores along the different

routes was inconclusive in the present study. Indeed Bolger et al.(2000) found that the abundance of some arthropods was enhancedby fragmentation.

The observation that the routes generated from the whole pixelMLC classifier, and from route A with the largest tree patches, wereless able to traverse the urban area, supports the assertions of oth-ers (eg. Benson and Mackenzie, 1995; Moody and Woodcock, 1995;O’Neill et al., 1996; Hay et al., 2001) that landscape parameters aredependent on grain size. Indeed, O’Neill et al. demonstrate that a10% change in land cover class due to change of resolution mayresult in an 80% change in habitat dominance within sample plots,and they recommend sampling grain sizes 2–5 times smaller thanthe spatial features of interest. In the present case, the coarser grainsizes of 20 m × 20 m (SPOT) and 4 m × 4 m (Ikonos) from the MLCclassifier were unable to identify a direct path across the urbanlandscape due to the apparent lack of habitat linkages at those res-olutions, compared with tree patches of ca. 20% of an Ikonos pixel(ca. 3.2 m2) detected by LSU.

7. Conclusion

In densely urbanized Hong Kong, tree patches of 400 m2 and16 m2 corresponding to the sizes of SPOT and Ikonos pixels, respec-tively, do not appear to provide viable connecting corridors forurban birds. Thus in our densely urbanized study area, wheregreen space is extremely fragmented (Fig. 2) sub-pixel analysis offine resolution images offers a more complete and relevant habi-tat inventory, and such detailed mapping at whole city level hasnot been undertaken previously. The fuzzy querying from the LSUsoft classifier produced different paths of least resistance for birdsacross the urban landscape, which were based on differing sizes ofmicro-scale habitats. The results suggest that in a highly urbanisedregion, connectivity is of great importance to most species. Themethod demonstrated here offers a means for testing the impor-tance of connectivity versus patch size, over a whole city, for grassand tree habitats either individually or combined, and permits thedesignation of relevant wildlife corridors. Moreover, the soft clas-sification approach described, facilitates upscaling to coarser grainsizes for enquiries at differing levels of landscape analysis (Blair,1996; Hay et al., 2001), by varying the fraction amounts within apixel, and/or by pixel resampling.

Acknowledgement

The study was supported by CERG grant no. PolyU5264/O8Efrom the Hong Kong government.

References

Adams, J.B., Smith, M.O., Johnson, P.E., 1986. Spectral mixture modeling: a new anal-ysis of rock and soil types at the Viking lander. J. Geophys. Res. 91, 8098–8122.

Ares, J.O., Dignani, J., Bertiller, M.B., 2007. Cost analysis of remotely sensed forag-ing paths in patchy landscapes with plant anti-herbivore defenses (Patagonia,Argentina). Landsc. Ecol. 22, 1291–1301.

Beissinger, S.R., Osborne, D.R., 1982. Effects of urbanization on avian communityorganization. Condor 84, 75–83.

Belisle, M., 2005. Measuring landscape connectivity: the challenge of behaviourallandscape ecology. Ecology 86 (8), 1988–1996.

Benson, B.J., Mackenzie, M.D., 1995. Effect of sensor spatial resolution on landscapestructure parameters. Landsc. Ecol. 10 (2), 113–120.

Berry, J.K., 1993. Beyond Mapping: Concepts, Algorithms and Issues in GIS. GIS WorldWiley.

Bino, G., Levin, N., Darawshi, S., Van der Hal, N., Reich-Solomono, A., Kark, S., 2008.Accurate prediction of bird species richness patterns in an urban environmentusing Landsat-derived NDVI and spectral unmixing. Int. J. Remote Sens. 29 (13),3675–3700.

Blair, R.B., 1996. Land use and avian species diversity along an urban gradient. Ecol.Appl. 6, 506–519.

Bolger, D.T., Suarez, A.V., Crooks, K.R., Morrison, S.A., Case, T.J., 2000. Arthropods inurban habitat fragments in southern California: area, age and edge effects. Ecol.Appl. 10 (4), 1230–1248.



Clergeau, P., Savard, J.P.L., Mennechex, G., Falardeau, G., 1998. Bird abundance anddiversity along an urban–rural gradient: a comparative study between two citieson different continents. Condor 100, 413–425.

DeGraaf, R.M., Geis, A.D., Healy, P.A., 1991. Bird population and habitat surveys inurban areas. Landsc. Urban Plann. 21, 181–188.

DeGraaf, R.M., Wentworth, J.M., 1991. Urban bird communities and habitats in NewEngland. Proc. Trans. North Am. Wildlife Natural Res. Conf. 46, 396–413.

Desrochers, A., Hannon, S.J., 1997. Gap crossing decisions by forest songbirds duringthe post-fledgling period. Conserv. Biol. 11 (5), 1204–1210.

Donnelly, R., Marzlufff, J.M., 2004. Importance of reserve size and landscape contextto urban bird conservation. Conserv. Biol. 18, 733–745.

Eastman, R., 2006. IDRISI Andes Tutorial. Clark Labs, Worcester, USA.Fahrig, L., 2003. Effects of habitat fragmentation on biodiversity. Annual review of

ecology. Evol. System. 34, 487–515.Fernandez-Zuricic, E., Jokimaki, J., 2001. A habitat island approach to conserving

birds in urban landscapes: cases studies from southern and northern Europe.Biodiv. Conser. 10, 2023–2043.

Hay, G.J., Marceau, D.J., Dube, P., Bouchard, A., 2001. A multiscale framework for land-scape analysis: object-specific analysis and upscaling. Landsc. Ecol. 16, 471–490.

Huste, A., Boulinier, T., 2007. Determinants of local extinction and turnover rates inurban bird communities. Ecol. Appl. 17 (1), 168–180.

Jim, C.Y., 2004. Impacts of intensive urbanisation on trees in Hong Kong. Environ.Conserv. 25 (2), 146–159.

Jokimaki, J., 1999. Occurrence of breeding bird species in urban parks: effects of parkstructure and broad scale variables. Urban Ecosyst. 2, 21–34.

Lock, N.Y., 2000. The ecology of urban birds in Hong Kong. Unpublished PhD thesis,University of Hong Kong, Hong Kong.

Lodwick, W., 2008. Fuzzy Surfaces in GIS and Geographical Analysis. CRC Press,Taylor & Francis.

Marzluff, J.M., 2005. Island biogeography for an urbanizing world: how extinctionand colonization may determine biological diversity in human dominated land-scapes. Urban Ecosyst. 8, 157–177.

Melles, S., Glenn, S., Martin, K., 2003. Urban bird diversity and landscape complexity:species–environment associations along a multiscale habitat gradient. Conserv.Ecol. 7 (1), art 5.

Molainen, A., Hanski, I., 2006. Connectivity and metapopulation dynamics in highlyfragmented landscapes. In: Crooks, K.R., Sanjayan, M. (Eds.), Connectivity Con-servation. Cambridge University Press, UK, pp. 44–71.

Moody, A., Woodcock, C.E., 1995. The influence of scale and the spatial characteristicsof landscapes on land cover mapping using remote sensing. Landsc. Ecol. 10 (6),363–379.

Nichol, J.E., Wong, M.S., 2007. Remote sensing of urban vegetation life form by spec-tral mixture analysis of high resolution Ikonos satellite images. Int. J. RemoteSens. 28 (5), 985–1000.

Neill, R.V., Hunsaker, C.T., Timmins, S.P., Jackson, B.L., Riitters, K.H., Wickham, J.D.,1996. Scale problems in reporting landscape pattern at the regional scale. Landsc.Ecol. 11 (3), 169–180.

Opdam, P., vanDorp, D., terBraak, C.J.F., 1984. The effect of isolation on the num-ber of woodland birds in small woods in The Netherlands. J. Biogeogr. 11,473–478.

Park, C.R., Lee, W.S., 2000. Relationship between species composition and areain breeding birds of urban woods in Seoul, Korea. Landsc. Urban Plann. 51,29–36.

Phinn, S., Stanford, M., Scarth, P., Murray, A.T., Shyy, P.T., 2002. Monitoringthe composition of urban environments based on the vegetation-impervioussurface-soil (VIS) model by subpixel analysis techniques. Int. J. Remote Sens. 23,4131–4153.

Porter, E., Forschner, B.R., Blair, R.B., 2001. Woody vegetation and canopy fragmen-tation along a forest-to-urban gradient. Urban Ecosyst. 5, 131–151.

Ridd, M., 1995. Exploring a V-I-S (vegetation-impervious surface-soil) model forurban ecosystem analysis through remote sensing: comparative anatomy forcities. Int. J. Remote Sens. 16, 2165–2185.

Savard, J.P.L., Clergeau, P., Mennechez, G., 2000. Biodiversity concepts and urbanecosystems. Landsc. Urban Plann. 48, 131–142.

Small, C., Lu, J.W.T., 2006. Estimation and vicarious validation of urban vegetationabundance by spectral mixture analysis. Rem. Sens. Environ. 100, 441–456.

Sodhi, N.S., Briffet, C., Kong, L., Yuen, B., 1999. Bird use of linear areas of a tropicalcity: implications for park connector design and management. Landsc. UrbanPlann. 45, 123–130.

Sukopp, H., Weiler, S., 1988. Biotope mapping and nature conservation strategies inurban areas of the Federal Republic of Germany. Landsc. Urban Plann. 15, 39–58.

Wee, Y.C., Corlett, R.T., 1986. The Forest and the City: Plant Life in Urban Singapore.Singapore University Press, 186 p.

Whitcomb, R.F., 1977. Island biogeography and ‘habitat islands’ of eastern forests.1. Introduction. Am. Birds 31, 3–5.

assessing avian habitat fragmentation in urban areas of

Documents